Automatic fiducial localization in brain images
Dingguo Chena, Jun Tana , Vipin Chaudharyb, Ishwar K Sethia
a
Dept of Computer Science and Engineering, Oakland University, USA
b
Institute for Scientific Computing , Wayne State University, USA
Abstract. Point-based registration is one of the most popular registration methods in
practice. During registration, we use external fiducial markers that are rigidly attached
through the skin to the skull. Determining the coordinates of the fiducials without
human selection is crucial for a fully automatic point-based image registration. In this
paper, an image processing technique for automatic fiducial detection is presented. The
technique, which is fast, automatic and accurate, has two major steps: Edge map
construction and curvature-based object detection. Experimental results are promising
compared to the manul selection by experts with no error for 56 CT images and 3%
error for 66 MR images.
Keywords: Fiducial point detection, edge map construction, curve-based corner detection
1. Introduction
Many methods have been used to register medical images. Point-based registration is
the most popular method in practice and used by many commercial navigation systems.
Point-based registration involves the determination of the coordinates of corresponding
points in different images (or physical space) and the estimation of the geometrical
transformation using these corresponding points. The points may be either intrinsic or
extrinsic. Intrinsic points are derived from naturally occurring features, e.g., anatomic
landmark points. Extrinsic points are derived from artificially applied markers, e.g.,
tubes containing copper sulfate. The estimation of fiducial points’ locations on brain
images is important to perform a successful point-based registration. As we all know,
manual selection is flawed and inefficient during the surgery. As a first step of
registration, there is some research already done about automatic fiducial detection in
the last few years. Matthew et al proposed an automatic technique for marker
localization using an intensity based search and classification [1]. Their algorithm
requires prior information on marker size and shape and could only be used for
cylindrical markers. Tan et al [10] proposed a template-based technique based on shape
information. Their technique is targeted for volumes. Lewis et al developed an
ultrasound module based method for detection and location of subcutaneous fiducials,
which involved additional hardware equipment [2]. Other techniques requiring
additional hardware include articulated arms [3, 4], optical triangulation systems [5],
and magnetic field digitizers [6]. In this paper we present a fiducial detection algorithm
using intensity information in image space, which is fully automatic and avoids using
additional hardware. The rest of this paper is organized as follows: in Section 2 we
introduce our algorithm including ratio-based segmentation, ray-based outline search,
corner detection and point clustering. The results are given in Section 3 followed by
discussion in Section 4.
2. Methods
The two steps, edge detection and curvature based 2D object detection are discussed in
detail in the following sections.
2.1 Edge Map Construction
As the first step, a ratio-based thresholding method [7] is used to segment the
foreground object from background in the original images. At the optimum threshold,
an obvious change of ratio between two region pixel numbers will occur. From the set
of potential threshold values thus located, a subset of final threshold values can be
automatically determined. The first thresholding value will be used for
foreground/background segmentation. The steps for performing segmentation using the
ratio curve are as follows:
• Compute the image histogram and smooth it using a Gaussian filter to
suppress small intensity variations
• Obtain the ratio curve. Assume the pixels in an image be represented by
L gray levels, [1, 2, ..., L], and let h j denote the number of pixels with
gray level j. Then the number of pixels at or below threshold
i
is N 1 = ∑ hj , and those above the threshold is N 2 =
j =1
L
∑ hj . Therefore the
j = i +1
ratio of the two numbers is R = N 1 N 2 , and the ratio curve exhibits a
variation of R.
• Compute the second derivative of ratio curve to obtain the normalized
rate curve. The normalized rate curve is also smoothened to flatten small
peaks and valleys. The locations of the resulting minima and maxima
determine the candidate threshold values.
The segmented result on an example image is depicted in Fig.1
(a) Original Image
(b) Segmented Object
Figure 1: Foreground/background Segmentation
In the segmented foreground, we employ morphological recursive erosion, dilation
and grayscale reconstruction techniques [8] to fuse small gaps and to eliminate
unwanted noises, often present in pathological images. Canny edge detection is used
next to find edges [8]. (Figure 2.a)
Among the numerous detected edges, the outline of brain is selected by the following
ray search procedure: 1) Rank the pixel number in each edge object and choose the top
half as our candidates; 2) Set four rays from corners to the image centre; 3) The first
edge to intersect any of the rays will be considered the outline of the brain. The
Gaussian smoothing filter is applied before the procedure to eliminate the small objects
and noises. (Figure 2.b)
Figure 6: Edge Detection
(a) Initial Edge Map
(b) Outline selected
Figure 2: Outline Selection
2.2 Curvature-based fiducial detection
Once the outline of the brain is identified, corners are detected within this outline.
Single-scale feature detection finds it hard to detect both fine and coarse features at the
same time. On the other hand, multi-scale feature detection [9] is able to solve this
problem. Using an adaptive local curvature threshold instead of a single global threshold
as in the original and enhanced CSS (Curvature Scale Space) methods the angles of
corner candidates are checked in a dynamic region of support for eliminating falsely
detected corners. Here the curvature is defined as [9]:
X& (u , σ )Y&&(u , σ ) − X&& (u , σ )Y& (u , σ )
k (u , σ ) =
( X& (u , σ ) 2 − Y& (u , σ ) 2 )1.5
Where X& (u , σ ) = x(u , σ ) ⊗ g& (u , σ ), X&& (u , σ ) = x(u , σ ) ⊗ g&&(u , σ ),
Y& (u , σ ) = y (u , σ ) ⊗ g& (u , σ ), Y&&(u , σ ) = y (u , σ ) ⊗ g&&(u , σ ), and ⊗ is the convolution
operator while g (u , σ ) denotes a Gaussian function with deviation σ .
We compute curvature at a fixed low scale for each contour to retain all true corners.
And all of the curvature local maxima are considered as corner candidates, including the
false corners. Among the initial candidates, adaptive local threshold and angel of corner
are used to remove the false corners and boundary noise. Corner selection result is
shown in Figure 3.
Figure 3: Corner Detection
Next, K-means is used to cluster the selected corner points into several classes based on
point positions. Polynomial curve fitting is then used to find the coefficients of a
polynomial P (x) of degree 3 that fits the data in each class. Thus, a fitted curve is
plotted in each group (Figure 4). The center point of each curve is selected and back
projected onto the brain contour. The resulting points are the center of the fiducial
markers (Figure 5).
Figure 4: Curve Fitting
Figure 5: Final Result
3. Experimental Results
The proposed technique was tested on 122 images from different patients and different
modalities: CT (512x512x56) and MR (256x256x66) (forward and reverse readout
gradient T1, PD and T2). All experiments were performed on 1.5GHZ CPU workstation
with 512MB Memory PC using Matlab 7.0 and Visual C++ 6.0.
The detected fiducial points were compared with manually selected points by experts.
If distance between selected points and ground truth is more than 0.5 mm, we classify it
as a wrong detection. For CT images, the error rate in zero using 56 images. For MR
images, the error rate is 3% (2 out of 66 images). The average distance is less than
0.4mm. The experimental results indicate that the algorithm is reliable and accurate.
4. Conclusion
In this paper, we proposed a novel automatic fiducial points detection algorithm usable
in different image modalities. This will be useful for automatic point based registration
[11]. The proposed technique was tested using CT and MR images with perfect
detection in CT images and 2 errors out of 66 MR images.
Reference
1)
2)
3)
4)
5)
6)
7)
8)
9)
10)
11)
Matthew Y. Wang, Calvin R. Maurer, Jr., J. Michael Fitzpatrick, and Robert J. Maciunas, “An
Automatic Technique for Finding and Localizing Externally Attached Markers in CT and MR
Volume Images of the Head”, IEEE Transaction on biomedical engineering, VOL. 43, NO. 6, June
1996 621
Thomas Lewis, Robert L. Galloway, Jr., Steven Schreiner, “An Ultrasonic Approach to Localization
of Fiducial Markers for Interactive, Image-Guided Neurosurgery—Part I: Principles.” IEEE
Transaction on biomedical engineering, VOL. 45, NO. 5, MAY 1998
L. Adams, W. Krybus, D. Meyer-Ebrecht, R. Rueger, J. M. Gilsbach, R. Moesges, and G.
Schloendorff, “Computer-assisted surgery,” IEEE Comput. Graphics Applicat., vol. 10, pp. 43-51,
1990.
R. L. Galloway, Jr., R. J. Maciunas, and C. A. Edwards, “Interactive image-guided neurosurgery,”
IEEE Trans. Biomed. Eng., vol. 39, pp.1226-1231, 1992.
R. L. Galloway, Jr., R. J. Maciunas, W. A. Bass, and D. Crouch, “An optical device for interactive,
image-guided neurosurgery,” in Proc. Annu. Int. Con& IEEE Eng. Med. Biol. Soc., 1993, vol. 15, pp.
954-955.
A. Kato, T. Yoshimine, T. Hayakawa, Y. Tomita, T. Ikeda, M. Mitomo, K. Harada, and H. Mogami,
“A frameless, armless navigational system for computer-assisted tomography,” J. Neurosurg., vol.
74, pp. 845-849, 1991.
Dingguo Chen, Ishwar K Sethi, Vipin Chaudhary, “Ratio-based automatic medical image
segmentation”, (Submitted)
Rafael C. Gonzales and Richard E. Woods , “Digital Image Processing”,: Prentice Hall, MA, 2002
X.C. He and N.H.C. Yung, "Curvature Scale Space Corner Detector with Adaptive Threshold and
Dynamic Region of Support", Proceedings of the 17th International Conference on Pattern
Recognition, 2:791-794, August 2004.
J. Tan, D. Chen, IK Sethi and V. Chaudhary, “A Template Based Technique for Automatic Detection
of Fiducial Markers in 3D Brain Images,” Proceedings 20th International Congress on Computer
Assisted Radiology and Surgery (CARS), Osaka, Japan, June 2006
G. Kaur, J. Tan, M. Alam, et al, "CASMIL: A comprehensive software/toolkit for Image-guided
Neurosurgeries", Neurosurgery edition of IJMRCAS (International Journal of Medical Robotics +
Computer Assisted Surgery) 2006 (To appear).